fix(llm): retry structured calls that return no result

[Thanatos9404]

Fixes #843

LangChain's tool-call parsers return None when a reply has no tool call. bind_structured_output already turns that into a StructuredOutputParseError for auto tool choice bindings (#529) and explicit function_calling (#691), so those are retried and end as llm_structured_response_invalid. Every other binding returns the None unchanged, including the default forced tool call on Bedrock (ChatBedrockConverse) and on the Anthropic API (ChatAnthropic). _invoke_batch then passes None to parse_response:

• _TP4Analyzer.parse_response raises NotImplementedError("TP4 requires a structured assessment response"). arun_batches_detailed re-raises NotImplementedError as a misconfiguration, so _check_tp4 records failed / llm_batch_failed after one attempt and the scan ends with execution_successful: false and exit code 2, as in the issue.
• The base parse_response, used by the three semantic analyzers, raises the same kind of NotImplementedError; the semantic node then reports the whole analyzer unavailable and drops the other batches' results.
• LLMMetaAnalyzer.parse_response fails with AttributeError, and the batch fails as llm_batch_failed without a retry.

This PR checks for a None result inside the existing try in _invoke_batch and _ainvoke_batch and raises StructuredOutputParseError there. A missing result now takes the same path as any other malformed structured response: three bounded retries, then a skipped batch with llm_structured_response_invalid. For TP4 the analyzer ends degraded, the report shows partial analysis, and execution_successful stays true. If a retry returns a valid assessment, TP4 completes normally.

The check sits in the base class rather than in TP4's parse_response because parse_response runs outside that try, and errors raised there deliberately propagate without a retry (test_custom_parser_validation_error_propagates_without_retry).

This is the follow-up suggested in the #691 review (#691 (comment)). #691 applied it to explicit function_calling and kept other bindings failing on purpose; this applies it to every structured binding.

On current main the issue's model, global.anthropic.claude-sonnet-5-5, is registered as rejecting forced tool calls (#806), so on Bedrock it already takes the auto tool choice path and is retried. Models that keep the forced tool call, such as the Bedrock default us.anthropic.claude-sonnet-4-6, still hit the crash on main. The 2.13.0 release branch (#833) leaves out #806 and has the same _invoke_batch code, so the issue's exact setup still exits 2 there. The llm_analyzer_base.py change applies cleanly to yashraj/release-v2.13.0 if a backport is wanted.

Tests

tests/nodes/test_llm_analyzer_base.py, in both TestRunBatches and TestARunBatches:

• test_missing_structured_response_recovers_on_retry: None, then a valid result. Two calls, one 0.5 s backoff, no failure.
• test_missing_structured_response_isolated_after_three_retries: four None results. The batch fails with ValidationError / llm_structured_response_invalid after the 0.5, 1 and 2 s backoffs, and the next batch still succeeds.

tests/test_mcp_tool_poisoning.py, in TestTP4Fallbacks:

• test_missing_structured_response_is_retried: None, then a valid assessment. TP4 ends completed after two calls.
• test_bedrock_forced_tool_call_answered_in_prose_is_retried: a real ChatBedrockConverse for BEDROCK_DEFAULT_MODEL (forced tool choice) whose _agenerate returns prose, so LangChain's parser returns None. TP4 makes four calls, ends degraded with llm_structured_response_invalid, and records no failed rows.
• test_tool_refusal_stays_incomplete (from fix: prefer tool output for OpenAI Opus 5 TP4 #691) pinned the old behaviour: one call, then failed. It now expects four calls, degraded, a skipped llm_structured_response_invalid row, and finalize_ledger returning execution_successful: true and is_complete: false.

All seven fail on main @ 2404212 (the base tests with NotImplementedError, the TP4 tests with one call where two or four are expected) and pass on this branch.

Checks

On Windows, Python 3.12.5, uv sync --frozen --all-extras, at 2404212 plus this commit:

• pytest -m "not integration and not provider" tests/: passes except about 40 Windows-only tests (environment size limit, os.geteuid, symlinks, paths and encodings, plus one timing-sensitive test) that fail the same way on unmodified main; none touch the LLM code.
• The LLM-related test files (tests/test_mcp_tool_poisoning.py, tests/nodes/test_llm_analyzer_base.py, tests/nodes/test_meta_analyzer.py, tests/nodes/test_meta_analyzer_fallback.py, tests/nodes/test_semantic_quality_policy.py, tests/nodes/analyzers/test_semantic_developer_intent.py, tests/nodes/analyzers/test_semantic_security_discovery.py, tests/test_semantic_runtime.py, tests/unit/test_llm_utils.py, tests/unit/test_tp4_openai_http.py, tests/unit/test_bedrock_provider.py): 770 passed.
• ruff check src/ tests/ and ruff format --check src/ tests/ (what make lint and make format-check run): pass.

I also ran the TP4 node offline against a real ChatBedrockConverse whose botocore client was stubbed with Stubber to answer in prose. For us.anthropic.claude-sonnet-4-6, main makes one Converse call and ends failed / NotImplementedError with execution_successful: false; this branch makes four calls and ends degraded / llm_structured_response_invalid with execution_successful: true, and a prose reply followed by a valid tool call completes after two calls. No live provider calls were made.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.